2558
5, MAY 2009
Training Symbol Based Coarse Timing
Synchronization in OFDM Systems
Ming (Matt) Ruan, Member, IEEE, Mark C. Reed, Senior Member, IEEE, and Zhenning Shi, Member, IEEE
Abstract—In orthogonal frequency division multiplexing
(OFDM) systems, coarse frame timing can be acquired from one
or more training symbols preceding every OFDM burst. The
existing literature studied the case where there was only one
training symbol consisting of identical segments. We generalize
the timing synchronization methods to take advantage of multiple
training symbols and only require the segments to be highly
correlated but not necessarily identical. We construct a series of
component timing metrics, one for each pair of the highly corre-
lated segments, and combine them linearly to minimize the false
alarm probability while keeping the asymptotic missed detection
probability to the same level as other techniques. The OFDM data
symbols in a downlink burst can have different power levels to
reach the users at different distances. We take that into account
and yield more realistic results than those in existing literature
which only considered the equal power case. The performance of
the proposed method is analyzed in three scenarios generalized
from the IEEE 802.11 and IEEE 802.16 standards. Numerical
results are presented to confirm the robustness of the proposed
method in various channel conditions.
Index Terms—Orthogonal frequency-division multiplexing
(OFDM), synchronization.
I. I NTRODUCTION
I N ORTHOGONAL frequency-division multiplexing
(OFDM) systems, coarse timing synchronization is
needed to indicate the start of a burst and determine the
Fast Fourier Transform (FFT) window for each symbol in
that burst. Schmidl and Cox [1] proposed to construct the
training symbol with two identical halves in the time domain
and estimate the beginning of OFDM symbols from the
auto-correlation of the received signal. Coulson refined that
algorithm in [2], and gave detailed performance analysis in
[3]. Minn et al. [4] proposed to use a training symbol with
more than two identical segments, and suggested to flip the
signs of the segments with certain pattern to give the timing
metric a steeper roll-off trajectory at the ideal timing position.
Shi and Serpedin analyzed Minn’s synchronization scheme
in [5] and proposed a more advanced timing metric based on
the maximum likelihood criterion.
The boundaries between the OFDM symbols are blurred
by multipath wireless channels, so the ideal start of the FFT
window at the receiver is usually difficult to locate in the
Manuscript received March 13, 2008; revised June 14, 2008 and October
23, 2008; accepted January 11, 2009. The associate editor coordinating the
review of this paper and approving it for publication was J. Coon.
The authors are with the National ICT Australia and affiliated with
the Australian National University (e-mail: {matt.ruan, mark.reed, zhen-
ning.shi}@nicta.com.au).
National ICT Australia is funded through the Australian Government’s
Backing Australia’s Ability initiative and in part through the Australian
Research Council.
Digital Object Identifier 10.1109/TWC.2009.080356
1536-1276/09$25.00 
Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO.
time domain. We know that, as long as the samples in the
FFT window belong to the same OFDM symbol, the inter-
symbol interference (ISI) and inter-carrier interference (ICI)
are avoided. Therefore, it suffices for the coarse timing esti-
mator to find an ISI-free FFT window of the training symbol.
The residual timing offset results in a linear phase shift on the
subcarriers, which can significantly degrade the performance
of pilot-based channel estimation and tracking algorithms [6].
Then, the refined joint channel and timing estimators [7]–
[9] are required for more accurate timing estimates using the
frequency domain signal.
In some practical OFDM systems, like the IEEE 802.11 [10]
and IEEE 802.16 [11], the training symbols are defined in the
frequency domain by using one out of every Mi subcarriers,
where i is the training symbol index. When Mi is not divisible
by the number of subcarriers, e.g., Mi = 3 for the downlink of
IEEE 802.16 OFDMA (WiMAX) systems, the training symbol
is made up of highly correlated but not identical segments in
the time domain. This scenario has not been investigated by
the aforementioned literature, and only a heuristic solution is
provided by Bhatt et al. [12].
This paper investigates the timing estimation methods based
on the training symbols generalized from those specified by
the standards [10], [11]. We propose to construct a series
of component timing metrics, one for each pair of highly
correlated segments in the training symbols. Then we linearly
combine them to minimize the false alarm probability under
the constraints on the correct detection performance. The
component timing metrics only require the correlation between
the segments to be high, so the proposed method is applicable
to generalized training symbols including those specified by
the IEEE 802.16 OFDMA (WiMAX) standard which have
three highly correlated but not identical segments. Moreover,
the proposed method takes advantage of multiple training
symbols for further performance improvement. It is also worth
noting that the data symbols in one OFDM burst can have
different power levels to reach the users at different distances.
We take that into account and yield more realistic results yet
found in the existing literature.
The performance of the proposed method is analyzed
in three scenarios generalized from the IEEE 802.11 and
IEEE 802.16 standards. When there is one training symbol
with two and four identical segments, the proposed method is
equivalent to that in [1] and [5] respectively. Our analytical
results show that using more identical segments in the training
symbol only reduces false alarms and has little impact on
the missed detection probability. This justifies our false alarm
probability based optimization criterion, and contradicts the
c 2009 IEEE
RUAN et al.: TRAINING SYMBOL BASED COARSE TIMING SYNCHRONIZATION IN OFDM SYSTEMS 2559

existing timing metric design approach which is based on the

detection probability. Simulations agree with our analytical I P Low
Add Pulse x(t) x[n]
results reasonably well and confirm the robustness of the F / Cyclic
Pass Sampling
Shaping
Filter
proposed method in various channel conditions. F S
Prefix Filter

This paper is organized as follows. We define our system T

model in Section II, and then derive the proposed algorithm
in Section III. The performance of the proposed method is Scope Display of the Traning Symbol
analyzed in Section IV for three different scenarios generalized x(t)
from the practical standards. Simulation results obtained with x[n]

realistic channel models are presented in Section V and a brief

summary concludes the paper in Section VI.
CP Period One Period Two Period Three
II. S YSTEM M ODEL
We study packet oriented OFDM systems where one or Fig. 1. Diagram of a generalized training symbol with N = 8, M1 = 3.
more training symbols precede every OFDM burst to facil-
itate timing, frequency and channel estimation. The training
symbols are defined in the frequency domain and only one where 0 is the true CFO normalized by the subcarrier spacing
out of every Mi subcarriers are used. More precisely, the time 1/(N Ts ); I is the number of training symbols; τ [l] is the
domain representation for the ith training symbol reads relative delay of the lth channel tap, h[l] is the complex
representation of the tap’s magnitude and phase; w[n] is
xi (t) = (s(t) − s(t − (N + Ng )Ts )) additive white Gaussian noise (AWGN) with variance σw 2
;
Np [i]
1  2π
the fixed phase offset between the transmitter and receiver
Xi [m]e−j N (ϑi +(m−1)Mi )( Ts −Ng )
t
·√ (1)
N m=1
is
absorbed
Lh inI the channel coefficients, and we define x̂[n] 
l=1 h[l] i=1 xi [n − (i − 1)(N + Ng ) − τ [l]]. In this paper,
where s(t) is the unit step function which equals 1 when t ≥ the signal to noise ratio for the ith training symbol is defined
0 and 0 otherwise; N is the number of subcarriers; Ts is as the ratio between the signal and noise power in the received
(1/N )-th of one symbol duration, which is also known as the OFDM symbol: SNRi  (σi2 /σw 2
), where
sampling period; Ng is the length of the cyclic prefix (CP) N +Ng −1

in terms of samples. For the ith training symbol, Np [i] is the 1
σi2  |x̂[n + (i − 1)(N + Ng )]|2
number of used subcarriers, ϑi is the index of the first used N
n=Ng
subcarrier, Xi [·] is the phase shift keying (PSK) modulated 
Lh
pseudo-random training sequence mapped to the subcarriers Np [1]
= |Xi [1]|2 |h[l]|2 . (4)
{ϑi , ϑi +Mi , · · · , ϑi +(Np [i]−1)Mi }. All the other subcarriers N
l=1
are not used. Using (1), for all t ∈ [i(N + Ng )Ts , (i + 1)(N + The second equality of (4) follows Parseval’s Theorem that the
Ng )Ts − Ts N/Mi ), we have
received signal’s power in the time domain should equal that
Np [i] in the frequency domain. It should be noted that the received
1 
xi (t + Ts N/Mi ) = √ Xi [m] signal within the CP may come from two OFDM symbols in
N m=1 a multipath channel, so we exclude it from the signal power
2π
· ej N (t/Ts +N/Mi −Ng )(ϑi +(m−1)Mi ) calculation.
2π
jM ϑi
=xi (t) e i . (2)
III. T HE M ETHOD
This indicates that the time domain waveform of the training In the time domain, the high autocorrelation between the
symbol is repeated every (N/Mi )Ts with a fixed phase offset segments of the training symbol distinguishes the preamble
of (2πϑi /Mi ). Figure 1 shows an example of such generalized from the noise and other OFDM symbols in the burst. This
training symbols with N = 8 and M1 = 3. The baseband means that timing can be estimated from any pair of correlated
signal xi (t) goes through the low-pass filter and is sampled segments. When multiple pairs exist in one or multiple training
at time interval Ts . We obtain a sequence of samples xi [n] = symbols, one would intuitively want to combine them for
xi (t)δ(t − nTs ) = xi (nTs ), where δ(·) is the Dirac delta enhanced performance. In this section, we first analyze the
function. If N is not an integer multiple of Mi , the period of statistical properties of the correlators and component timing
x(t) is not equal to an integer number of samples, and the metrics, then derive the coefficients to combine them in a lin-
repetition structure is not preserved in the sample sequence ear manner such that the false alarm probability is minimized
x[n]. Figure 1 has illustrated this fact. The samples of the sine under the constraints on missed detection performance. A brief
wave x(t) are not periodical although the original waveform summary concludes the section.
is.
After low-pass filtering and decimation, the baseband re- A. Correlators
ceived signal can be represented as
Since the high autocorrelation distinguishes the training
2π
r[n] = ej N n0 (x̂[n] + w[n]) (3) symbol from others, we follow the traditional works [1]–[5]

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
2560 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 5, MAY 2009

and base our timing estimator on the correlators defined as Because the knowledge of signal power σi2 is not available
−d−1
N to the receiver, the ideal normalization is not practical. An
R[n, d]  r∗ [n + k]r[n + k + d] (5) approximation to that is to use R[n̂i , 0] as an estimate of N σi2 ,
k=0
and we define the component timing metric as
where (·)∗ denotes complex conjugate. The correlator given by N 1 |R[n, di,k ]|
Tc [n, di,k ]  · · . (12)
(5) can be easily implemented on hardware using the integrate- N − di,k |ρi (di,k )| R[n, 0]
and-dump algorithm [3]. At perfect timing position n̂i , using (6), we have
Denote n̂i as the ideal start of the FFT window at the N 1 |R[n̂i , d]|
receiver for the ith training symbol. In Appendix A, we show Tc [n̂i , d] = ·
N − d |ρi (d)| R[n̂i , 0]
that when the signal is stronger than the noise and d is close
to a multiple of N/Mi , N 1 (N − d)|ρi (d)|σi2
≈ ·
2π   N − d |ρi (d)| N (σi2 + σw2)
R[n̂i , d] ≈ ej N 0 d (N − d)σi2 φi (d) ρi (d) + Θ(n̂i , d) (6) σ2
= 2 i 2. (13)
where σi + σw
−d−1
N This indicates that at high SNR the timing metric is very close
Θ(n̂i , d)  w∗ [n̂i + k]x̂[n̂i + k + d] to 1. We also note that the timing metrics in the references
k=0 [1], [2], [4] are equivalent to the component timing metric
N−1 Tc [n, (N/Mi )] when Mi is divisible by N .
+ w[n̂i + k]x̂∗ [n̂i + k − d] (7) Denote n̄i as a timing position that is far away from the
k=d ideal one n̂i , i.e., none of the samples from n̄i to n̄i + N − 1
and belong to the ith training symbol. In Appendix B, we show
 that in AWGN channels and for both equal and unequal power
1, d is multiple of (N/Mi );
ρi (d)  sin( N
π
Mi d Np [i])
cases, Tc [n̄i , di,k ] can be well approximated by a Rayleigh dis-
Np [i] sin( N
π
Mi d)
, others. tributed random variable with cumulative distribution function
(8) (cdf)
 2π  λ 2
− λ2
ej N ϑi d
, d is multiple of (N/Mi ); P (Tc [n̄i , di,k ] < λ) ≈ 2
f (x|σ̄i,k ) dx = 1 − e
2σ̄
(14)
φi (d)  π d (Mi (Np [i]−1)+2ϑi )
i,k

e j N , others. 0
x2
(9) −
2σ̄2
where 2
f (x|σ̄i,k )
= σ̄2 e x 2
and σ̄i,k = 12 ρ2 (d1i,k ) N −d
i,k 1
.
We can see that only the first term in the bracket of the right i,k i i,k
Our simulation shows that, in multipath channels where the
hand side of (6) represents the useful signal, whose magnitude
channel delay spread is much shorter than the OFDM symbol
is given by (N − d)|ρi (d)|. Because |ρi (d)| decays quickly as
duration, (14) is also a good approximation to the true statistics
d moves away from a multiple of N/Mi , we propose to choose
of the component timing metric.
di,k  [k N/Mi ], (10)
C. Combining the Timing Metrics
where k ∈ [1, Mi −1], and [·] denotes the function that outputs
Multiple component timing metrics can be constructed for
the integer closest to its argument. The existing papers [1]–[5]
have only studied the case where N/Mi is an integer, while the training symbols. Although each one of them gives a
this paper investigates a more general scenario that covers the timing estimate, it is possible to combine them for better
generic WiMAX training symbols where Mi = 3. It is worth performance. We propose to linearly combine the component
noting that when d = 0, E {R[n̂i , 0]} = N (σi2 + σw 2
) gives timing metrics to minimize the false alarm probability while
keeping the asymptotic missed detection probability low. The
the signal plus noise energy.
combined timing metric is defined as

B. Component Timing Metrics 

I 
Ki
T [n]  ωi,k Tc [n + (i − 1)(N + Ng ), di,k ] (15)
A timing metric measures the likelihood of a given timing i=1 k=1
position to be the start of a training symbol. We know that where Ki = Mi − 1 is the number of component metrics con-
the correlators can detect the highly correlated segments in structed for training symbol i, and {ωi,k } are the coefficients
a training symbol, so each one of them can be a timing that need to be determined according to two criteria:
metric. However, as revealed by (6), the output of a corre- • In a noiseless channel, at the ideal timing position n̂1 ,
lator fluctuates with the received signal strength. To avoid T [n̂1 ] ≈ 1.
the adverse impact of power fluctuation, we normalize the • At a wrong timing position n̄1 , the probability for T [n̄1 ]
correlator so that at the ideal timing position n̂i the expectation to reach the given threshold λ should be minimized.
of the correlation equals 1. Ideally, the normalization can be
To meet the first criterion, from (13) and (15), it requires that
performed as

I 
Ki
|R[n̂i , di,k ]| N 1 |R[n̂i , di,k ]| ωi,k = 1. (16)
= · · . (11)
E {|R[n̂i , di,k ]|} N − di,k |ρi (di,k )| N σi2 i=1 k=1

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
RUAN et al.: TRAINING SYMBOL BASED COARSE TIMING SYNCHRONIZATION IN OFDM SYSTEMS 2561

Shi and Serpedin’s

Under this constraint, we compute the coefficients {ωi,k } estimates Our estimate
satisfying the second criterion as follows.
Threshold 2
At a wrong timing position n̄1 , a false alarm occurs if the
combined timing metric reaches the threshold λ. As shown in Threshold 1
Appendix C, the false alarm probability is given by

√ K−1 K−1 − 2 I Ki ω2 σ̄2

λ2
Correct Timing Window
P (T [n̄] ≥ λ) ≈ ( 2π) λ e i=1 k=1 i,k i,k
 Ki 2 2
I Fig. 2. Comparison of coarse timing estimates from a typical timing metric.
i=1 k=1 ωi,k σ̄i,k
·
(17)
I Ki 2 2 K− 12
i=1 ω σ̄
k=1 i,k i,k IV. P ERFORMANCE A NALYSIS AND C ASE S TUDY
I To provide more insights into the proposed algorithm, we
where K = i=1 Ki . For high thresholds, the false alarm
probability is determined by the denominator of the exponent analyze the performance in three special cases generalized
from the IEEE 802.11 [10] and IEEE 802.16 [11] standards.

I 
Ki
In Case A, we compare our estimator to those conventional
2 2
 ωi,k σ̄i,k . (18) methods [1], [5] and shed new light on them; in Case B, we
i=1 k=1
investigate the gain of combining the second training symbol
Under the constraint in (16), we can show the coefficients into the timing metric; in Case C, we analyze the performance
{ωi,k } that minimize  are given by of the proposed method for generic WiMAX training symbols.
ρ2i (di,k )(N − di,k )
ωi,k = I  Ki . (19) A. Case A: One training symbol, N is multiple of M1
2
i=1 m=1 ρi (di,m )(N − di,m )
This case generalizes the uplink training symbol in
And, it follows that IEEE 802.16 [11] OFDM systems where the data OFDM
1 symbols are placed after the training symbol which is QPSK
 = Ip Ki . (20)
2 − di,n ) modulated and M1 = 4. From (19), we can compute
i−1 n=1 ρi (di,n )(N
(N − kN/M1 ) 2(M1 − k)
This means that the component metrics must have large ω1,k = M1 −1 = . (21)
|ρi (di,k )| values to make significant contribution to the com- m=1 (N − mN/M 1 ) M 1 (M1 − 1)

bined timing metric. This justifies our selection of {di,k } in The combined timing metric becomes
(10).  
M1 −1 N−i MN1 ∗ N 

i=1  k=0 r [n + k]r[n + k + i M1 ]
2
D. Summary T [n] = N−1 .
M1 − 1 2 |r[n + k]|
k=0
We briefly summarize the proposed coarse timing estimator (22)
as follows. Although derived from different optimization criteria, the
proposed timing metric coincides with those in [1], [5] respec-
1) Construct (Mi −1) component timing metrics Tc [n, di,k ]
tively for M1 = 2 and M1 = 4 cases. It is easy to compute
for each training symbol using (12) where {di,k } are
given by (10). 
M1 −1
2 2 2
2) Compute the combined timing metric T [n] using (15) = ωi,k σ̄i,k = . (23)
N (M1 − 1)
where {ωi,k } are given by (19). k=1

3) Once the combined timing metric T [n] reaches the This indicates that the false alarm probability decreases as the
detection threshold λ, the position corresponding to the number of identical segments increases.
maximum of the timing metric within the following N Next, we analyze the missed detection probability. Define
OFDM samples is the coarse timing estimate.
ZA [n̂1 ]  RA [n̂1 ] − λ R[n̂1 , 0] (24)
It should be noted that choosing the maximum of the timing  M1 −1
metric within certain window avoids the dependency of coarse where RA [n̂1 ]  M12−1 k=1 |R[n̂1 , d1,k ]|. As shown in
timing estimates on the threshold setting. Shi and Serpedin [5] Appendix D,
assumed the coarse timing estimate to be where the threshold
E {ZA [n̂1 ]} = N σ12 (1 − λ) − N σw
2
λ (25)
is first reached. Figure 2 shows an example to illustrate the
difference between their approach and ours. It is shown that V ar {ZA [n̂1 ]} ≈ 2N σ12 σw
2
(1 − λ) .2
(26)
our estimates are independent of the threshold settings and Following [3] and [5], we approximate ZA [n̂1 ] by a Gaussian
must fall in the correct timing window for any threshold random variable, and the missed detection probability is given
lower than the peak. However, Shi and Serpedin’s estimates by
are determined by the threshold settings, which have to be
precisely set close to the peak to obtain correct estimates. P (T [n̂1 ] < λ) = P (ZA [n̂1 ] < 0)
√   
In practical OFDM systems where the timing metrics change 1 N σ12 λ 2
σw
dramatically from burst to burst due to the channel variation, ≈ erfc − (27)
2 2 2
σw 1 − λ σ12
our approach gives more robust performance.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
2562 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 5, MAY 2009

0
10
standards. In the former, N = 128, M1 = 8, M2 = 2; in
Sim. M =2
1
the latter, N = 256, M1 = 4, M2 = 2. From (19), we can
−1
10 Sim. M1=4 calculate
Sim. M1=8 N − kN/Mi
−2
10 Ana. ωi,k = M1 −1 M2 −1
Missed i=1 (N − iN/M1 ) + i=1 (N − iN/M2 )
Error Probability

Detection
−3
10 2(Mi − k)
= . (28)
Mi (M1 + M2 − 2)
−4
10
False Alarm
6dB
The false alarm probability decays exponentially with the
10dB
−5
10
inverse of
2 Mi −1
2 2 2 1
−6
10 = ωi,k σ̄i,k = . (29)
i=1
N M1 + M2 − 2
k=1
−7
10
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Comparing (29) to that of Case A in (23), we can see the
Detection Threshold (λ)
benefit of combining the long preamble is negligible when
the number of identical segments in the short training symbol
Fig. 3. Performance in AWGN channel (Case A). is large, i.e., in the IEEE 802.11 case. However, for smaller
M1 as in the IEEE 802.16 case, the combined metric gives
 +∞ 2 much fewer false alarms.
where erfc(z)  √2π z e−t dt. This indicates that the
Similar to the approach we used in Case A, define
missed detection
 probability
 monotonously decreases to zero
as the SNR σ12 /σw 2
increases. And, to limit the missed ZB [n̂1 ]  RB,1 [n̂1 ]R[n̂2 , 0] + RB,2 [n̂2 ]R[n̂1 , 0]
σ12
detection probability within 50%, it requires λ ≤ σ2 +σ 2. − λR[n̂1 , 0]R[n̂2 , 0] (30)
This upper bound of λ varies from 0.5 to 0.9 as the SNR
w 1
2
 M −1
where RB,i [n̂i ]  M1 +M 2 −2 k=1 |R[n̂i , di,k ]|. In Ap-
i

increases from 0dB to 10dB, which indicates that the missed pendix E, we show that
detection probability is highly sensitive to the noise level, and
optimizing the threshold for a large range of SNR is virtually E {ZB [n̂1 ]} = N 2 σ12 (σ12 + σw
2
) − λN 2 (σ12 + σw
2 2
) (31)
impossible. Another interesting observation from (27) is that V ar {ZB [n̂1 ]} ≈ 4N 3 σ16 σw
2
(1 − λ)2 . (32)
the missed-detection probability does not depend on M1 . This
Thus,
means that using more identical segments in the training
symbol will not lower the missed detection probability. In P (T [n̂1 ] < λ) = P (ZB [n̂1 ] < 0)
other words, Schmidl’s estimator [1] cannot be outperformed √   
1 N σ12 λ 2
σw
in terms of missed detection probability by more complex ≈ erfc √ − . (33)
algorithms including Shi and Serpedin’s method [5]. 2 2 2 2
σw 1 − λ σ12
We numerically evaluate the missed detection and false Compared to (27),
alarm probabilities of Case A for an unequal power OFDM • The 50% missed detection probability corresponds to λ =
system with 512 subcarriers in an AWGN channel. 5% of the σ12 /(σ12 + σw
2
), which is same as that of Case A.
subcarriers at each end of the spectrum are unused, and the • Combining the timing metric of the second symbol
CFO is modeled by a random variable uniformly distributed slightly increases the missed detections at high SNR.
in [−10, 10] subcarrier spacing. When testing the false alarm • Because the missed detection probability given by (33)
probability, the relative power level of each OFDM symbol is not a function of M1 or M2 , using more identical
is independently selected from {−12dB, −9dB, · · · , +9dB} segments in the training symbols does not reduce missed
with equal probability and the noise power is fixed to −10dB. detections.
The modulation scheme for each data symbol is independently We verify the analytical results using the same simula-
selected from QPSK, 16-QAM or 64-QAM also with equal tion environment as that of Case A and plot the results in
probability. When testing the missed detection probability, the Figure 4. The analytical and simulation results for Case B
training symbol is QPSK modulated and randomly generated are represented by the solid lines and the various markers
for each experiment, and we change the power of the training respectively. The dotted lines are the analytical results in the
symbol to validate our analytical results. In Figure 3, the scenario where only the short training symbol is utilized. It is
solid lines represent the analytical results given by (17) and shown that our analysis agrees with the simulations reasonably
(27), and the various markers represent the simulation results well. The figure suggests that using both training symbols
averaged for at least 105 independent OFDM symbols. The can considerably reduce the false alarm probability for the
figure shows that our analysis matches the simulation results M1 = 4 case, but the improvement is marginal for the M1 = 8
very well. case. The missed detection curves of Case A and B are very
close to each other under the given SNRs, and all of them
B. Case B: Two training symbols, N is multiple of both M1 exhibit very steep rising edges. This means that combining
and M2 two training symbols does not have practical impact on the
This case generalizes the short and long training symbol detection performance, which is largely determined by the
schemes specified in the IEEE 802.11 and IEEE 802.16 SNR.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
RUAN et al.: TRAINING SYMBOL BASED COARSE TIMING SYNCHRONIZATION IN OFDM SYSTEMS 2563

0 0
10 10
Sim. M1=4, M2=2 Sim. M1=3
−1 −1 Sim. M1=5
10 Sim. M =8, M =2 10
1 2
Ana. Ana. M1=3
−2 Case A Reference (same M1) −2
Ana. M =5
10 10 1
Missed Case A Reference Missed
Error Probability

Error Probability
−3
Detection −3
Detection
10 10

−4 −4
10 10
False Alarm False Alarm
6dB 10dB 6dB 10dB
−5 −5
10 10

−6 −6
10 10

−7 −7
10 10
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Detection Threshold (λ) Detection Threshold (λ)

Fig. 4. Performance in AWGN channel (Case B). Fig. 5. Performance in AWGN channel (Case C).

C. Case C: One training symbol, N is not multiple of M1 simulation results. Also, we use dotted lines to represent the
This case generalizes the IEEE 802.16 OFDMA (WiMAX) analytical missed detection probabilities of Case A under the
downlink channel where M1 = 3. In this case, given SNRs. It is shown that our analysis agrees with the
simulations reasonably well. The figure suggests that using
(N − d1,k )ρ21 (d1,k ) more highly correlated segments in the training symbols re-
ω1,k = M1 −1
2
m=1 ρ1 (d1,m )(N − d1,m ) duces false alarms but has little impact on the missed detection
2(N − d1,k ) ρ21 (d1,k ) probability, which is largely determined by the SNR. These
= M1 −1 2 (34) results coincide with our findings in the former two cases.
N m=1 ρ1 (d1,m )

where the last equality follows the fact that |ρi (d)| = V. S IMULATION R ESULTS
|ρi (N − d)|. The combined metric’s false alarm probability
In this section, we present simulation results in realistic
is determined by
wireless communication scenarios. An IEEE 802.16 [11]

M1 −1 (WiMAX) system is modeled in the simulations using two
2 2 2 1
= ω1,k σ̄1,k = M1 −1 2 . (35) kinds of training symbols specified for the OFDM and
N k=1 ρ1 (d1,k )
k=1 OFDMA physical layers respectively. The OFDM physical
Define layer has N = 256 subcarriers, and employs two training
ZC [n̂1 ]  RC [n̂1 ] − λR[n̂1 , 0] (36) symbols with M1 = 4 and M2 = 2. The OFDMA physical
 layer has N = 512 subcarriers and uses only one training
where RC [n̂1 ]  M1 −12 2
K1
k=1 |ρ1 (d1,k )R[n̂1 , d1,k ]|.
ρ1 (d1,k ) symbol with M1 = 3. The carrier frequency is set to 3.5GHz,
k=1
In Appendix D, we show that cyclic prefix is 1/8 of one symbol duration. The true CFO is
modeled by a uniformly distributed random variable within
E {ZC [n̂1 ]} = N σ12 − N λ(σ12 + σw
2
) (37) ±20 subcarrier spacing. Each OFDM burst lasts for 5ms,
and containing 47 OFDM symbols plus an idle period between
the bursts. Every point in the figures is an average of at least
V ar {ZC [n̂1 ]} ≈ 2λ2 N σ12 σw
2
− 4λN σ12 σw
2
3 × 104 independent experiments, each of which contains at
2 M −1 M −1 least one complete OFDM burst, but the starting position of
2  1 1

+ M1 −1 ρ(d1,k1 )ρ(d1,k2 )
2
k=1 |ρ(d1,k )| k1 =1 k2 =1
· ((N − max(d1,k1 , d1,k2 ))ρ(d1,k2 − d1,k1 )
+ max(N − d1,k1 − d1,k2 , 0)ρ(d1,k2 + d1,k1 )).
From (37), we can tell that the threshold corresponding to
50% missed detection probability of Case C is the same as
those of Case A and B. Considering the steepness of the
missed detection curves, we expect the detection performance
of Case C to be similar to that in the former two cases.
We verify the analytical results using the same simulation
environment as that of Case A. The results are plotted in
Figure 5 where we use solid and dashed lines to represent
the analytical results, and various markers to represent the
observation is randomly selected. The training symbols have
equal power, which is 9dB higher than the average power
level of data symbols. Each data symbol’s relative power level
(38) is independently selected from {−12dB, −9dB, · · · , +9dB}
with equal probability, and the modulation scheme is randomly
selected from QPSK, 16-QAM or 64-QAM also with equal
probability. The SNR is defined as the ratio between the
average power of the signal and noise in every received OFDM
sample of the training symbol.
Two channel models are used in the simulations, namely
“CH-A” and “CH-B”. CH-A is a stationary wireless com-
munication channel known as SUI-3 [13]. The channel has
3 taps with relative delays {0, 0.4μs, 0.9μs} and power
{0dB, −5dB, −10dB}. The magnitude of the first channel

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
2564 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 5, MAY 2009

tap follows Rice distribution, and those of the other two are 10
0

Rayleigh distributed. CH-B follows the power delay profile of

the Vehicular-A channel in [14], which has 6 taps with relative
delays {0, 0.31μs, 0.71μs, 1.09μs, 1.73μs, 2.51μs} and power

False Alarm Probability per Frame

−1
10
{0dB, −1dB, −9dB, −10dB, −15dB, −20dB}. The Doppler
frequency of CH-A is only 0.4Hz, so within each OFDM
symbol, the channel almost remains the same. We use Jakes’ 10
−2

model [15] to emulate the Rayleigh fading on each tap of CH- 0dB, CH−A, Prop.
0dB, CH−A, Conv.
B where the vehicle speed is set to 120km/h and the maximum 0dB, CH−B, Prop.
0dB, CH−B, Conv.
Doppler frequency is given by 10
−3

20dB, CH−A, Prop.

20dB, CH−A, Conv.
120km/h 3.5 × 109 Hz 20dB, CH−B, Prop.
fd = ≈ 388.9Hz. (39) 20dB, CH−B, Conv.
3 × 108 m/s 10
−4

0.1 0.15 0.2 0.25 0.3

In the simulations, each OFDM symbol lasts for 91μs, which Threshold (λ)

gives N fd Ts = 0.0354. We move the channel taps to the near- (a) OFDM physical layer (M1 = 4, M2 = 2)
est sample-spaced position to simplify the channel emulation.
In the simulations, we take the maximum of the timing 0
10
metric within the window [n̂i − N, n̂i + N ] as the timing
estimate. As we explained earlier, our coarse timing estimates
do not depend on the threshold settings, and for all the

False Alarm Probability per Frame

−1
10
thresholds that allow the detection of the OFDM burst, the
timing estimates will be the same.
For the OFDM physical layer, we refer to the method of [5] −2
10
as the conventional method, which only uses the short training 0dB, CH−A, Prop.
0dB, CH−A, Conv.
symbol for timing estimation. In Figure 6(a), 7(a) and 8(a), 0dB, CH−B, Prop.
we compare that to the proposed method which can take −3
10
0dB, CH−B, Conv.
20dB, CH−A, Prop.
advantage of multiple training symbols. For the OFDMA 20dB, CH−A, Conv.
physical layer, we refer to the method of [12] as the con- 20dB, CH−B, Prop.
20dB, CH−B, Conv.
ventional method, which only uses the first component timing −4
10
0.1 0.15 0.2 0.25 0.3
metric for timing estimation. In Figure 6(b), 7(b) and 8(b), Threshold (λ)

we compare that to the proposed method which employs two

(b) OFDMA physical layer (L1 = 3)
timing metrics combined for lower false alarm rate.In both
OFDM and OFDMA cases, the proposed methods require one Fig. 6. False alarm probability per frame.
more component timing metric than the conventional methods.

However, the curves of different SNRs differ dramatically

A. False Alarm Probability
Figure 6(a) and 6(b) show the false alarm probabilities as a
function of the threshold λ for OFDM and OFDMA physical
layers respectively. It is shown that the proposed method
achieves much better performance than the conventional meth-
ods [5], [12] in all the cases. The false alarm probability
at 0dB is lower than that at 20dB. This can be explained
by the fact that higher SNR causes larger power variation
within one OFDM symbol, which violates our assumptions
when deriving the variance of the component timing metrics in
Appendix B. For the same reason, the time-selectivity in CH-
B degrades the performance of the combined timing metric.
However, the figure shows that the conventional methods also
suffer from the power fluctuation, and the advantage of the
proposed method in terms of missed detection performance is
quite evident.
B. Missed Detection Probability
The missed detection probabilities are plotted in Fig-
ure 7(a) and 7(b) respectively for OFDM and OFDMA
physical layers. The figures show that virtually no difference
between the missed detection performance of the proposed
method and that of the conventional methods [5], [12].
with each other. This confirms our analytical results that
the missed detection probability is highly sensitive to SNR,
which fluctuates with the channel condition in a wide range.
Therefore, optimizing the detection performance at certain
SNR cannot give a timing metric with stable performance
in various channel conditions. Nevertheless, for any given
false alarm probability, our method allows the use of a lower
threshold, which gives a better chance to detect the right
timing position.
C. Distribution of Timing Offsets
As explained in the introduction, we define the correct
timing positions to be those that incur little or no ISI in
the FFT windows. We take the middle point of CP as the
reference timing position, and compute the probability for the
coarse timing estimates falling out of certain distance from the
reference point. In the OFDM and OFDMA cases, the lengths
of CP are 32 and 64 samples respectively, which means that
the correct estimates need to be within ±16 and ±32 samples
from the reference point for the OFDM and OFDMA cases
respectively.
Using the criterion described above, in Figure 8(a) and 8(b),
we see about 90% of the estimates fall in the window of

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
RUAN et al.: TRAINING SYMBOL BASED COARSE TIMING SYNCHRONIZATION IN OFDM SYSTEMS 2565

0 0
10 10
10dB, CH−A, Prop.
10dB, CH−A, Conv.
10dB, CH−B, Prop.

Probability of Falling out of the Distance

Missed Detection Probability per Frame

−1 10dB, CH−B, Conv. −1

10 10
20dB, CH−A, Prop.
20dB, CH−A, Conv.
20dB, CH−B, Prop.
20dB, CH−B, Conv.
−2 −2
10 10
10dB, CH−A, Prop.
10dB, CH−A, Conv.
10dB, CH−B, Prop.
−3 −3 10dB, CH−B, Conv.
10 10
20dB, CH−A, Prop.
20dB, CH−A, Conv.
20dB, CH−B, Prop.
20dB, CH−B, Conv.
−4 −4
10 10
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 5 10 15 20 25 30
Threshold (λ) Distance from the Middle of the CP (samples)

(a) OFDM physical layer (M1 = 4, M2 = 2) (a) OFDM physical layer (M1 = 4, M2 = 2)

0 0
10 10
10dB, CH−A, Prop.
10dB, CH−A, Conv.
10dB, CH−B, Prop.

Probability of Falling out of the Distance

Missed Detection Probability per Frame

−1 10dB, CH−B, Conv. −1

10 10
20dB, CH−A, Prop.
20dB, CH−A, Conv.
20dB, CH−B, Prop.
20dB, CH−B, Conv.
−2 −2
10 10
10dB, CH−A, Prop.
10dB, CH−A, Conv.
10dB, CH−B, Prop.
−3 −3 10dB, CH−B, Conv.
10 10
20dB, CH−A, Prop.
20dB, CH−A, Conv.
20dB, CH−B, Prop.
20dB, CH−B, Conv.
−4 −4
10 10
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 10 20 30 40 50 60
Threshold (λ) Distance from the Middle of the CP (samples)

(b) OFDMA physical layer (L1 = 3) (b) OFDMA physical layer (L1 = 3)

Fig. 7. Missed detection probability per frame. Fig. 8. Histogram of timing estimation error.

correct estimates. Very steep falling edges are observed exactly A PPENDIX A
at Ng /2. This indicates that the probability for large timing A PPROXIMATION TO R[n̂i , d]
offsets or significant ISI would decay quickly out of the From (5), we can expand R[n̂i , d] as
window of correct estimates. It is observed that the estimates N −d−1
of the proposed method are distributed in a narrower region 
j 2π
around the reference timing position than that of the conven- R[n̂i , d] = e N d 0
x̂∗ [n̂i + k]x̂[n̂i + k + d]
tional methods [5], [12]. This suggests superior performance k=0

−d−1
N
of the proposed algorithm. For both proposed and conventional ∗
+ w [n̂i + k]w[n̂i + k + d] + Θ(n̂i , d)
estimators, the coarse timing estimates are evenly distributed
k=0
in the window of correct estimates, so the refined timing (40)
estimation methods [7]–[9] are needed to determine the ideal
timing positions. where Θ(n̂i , d) is defined in (7). The second term in the
bracket of (40) is the summation of products of uncorrelated
VI. C ONCLUSIONS noise terms, its contribution to the value of R[n̂i , d] is much
smaller than the other terms when the signal is not weaker
A universal timing metric combining algorithm is proposed
than the noise. The first term in the bracket is the summation
in this paper to provide robust timing estimation performance
of products of signal terms, which can be expanded as
in various channel conditions. The proposed method can take
advantage of multiple training symbols and work in the sce- x̂∗ [n̂i + k]x̂[n̂i + k + d]
nario where the training symbols consist of highly correlated 
Lh 
Lh
but not identical segments. Our analysis shows that using more = h ∗
[l1 ]x∗i [k − τ [l1 ]] h[l2 ]xi [k + d − τ [l2 ]]
correlated segments in the training symbol reduces the false l1 =1 l2 =1
alarms but has little impact on the missed detections. The 
Lh
simulation results under various channel conditions confirm = |h[l1 ]|2 x∗i [k − τ [l1 ]]xi [k + d − τ [l1 ]]
the robustness of the proposed method. l1 =1

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
2566 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 5, MAY 2009


Lh 
(a) d ≤ N/2 Symb A Symb B
+ h∗ [l1 ]h[l2 ]x∗i [k − τ [l1 ]]xi [k + d − τ [l2 ]]. (41)
l1 =1 l2 =l1

The first summation in (41) contains the products of samples

with fixed distance d, which can be evaluated as
Np [i]
d N −d Ng
∗ 1  j 2π (ϑi +(m2 −1)Mi )d
xi [k] xi [k + d] = e N
N m =1 (b) d > N/2
⎛ 2
⎞ lb

· ⎝|Xi [m2 ]|2 + Xi∗ [m1 ]Xi [m2 ]ej N (m2 −m1 )Mi k ⎠ .
2π

m1 =m2
(42)
d
Because Xi [·] is a pseudo-random PSK modulated training
sequence, the summation of the second term in the bracket Fig. 9. Diagram of the component timing metric in the unequal power case.
of (42) is a summation of phase rotating terms with equal
magnitude, the value is close to 0. However, with properly
selected d, the first term accumulates over the summation and The samples in the correlation window, i.e., from sample n̄i
will dominate the value of the equation. It follows that to (n̄i +N −1), can belong to two OFDM symbols of different
Np [i]
power levels. Assume the boundary of the two symbols is at
1  j 2π (ϑi +(m2 −1)Mi ) d sample (n̄i +(N +Ng −lb )). From Figure 9, we can see that the
xi [k]∗ xi [k + d] ≈ |Xi [1]|2 e N
N m =1 number of samples belonging to the latter OFDM symbol is
2
given by lB = max(0, lb −Ng ), and that of the former symbol
Np
= |Xi [1]|2 ρi (d) φi (d) (43) is lA = (N − max(0, lb − Ng )). Denote the variance of the
N 2
samples of the former symbol as σA , and that of the latter as
where ρi (d) and φi (d) are defined in (8), and (9) respectively, 2
η·σA . It is worth mentioning that substantial power fluctuation
and due to the PSK modulation, |Xi [m2 ]|2 = |Xi [1]|2 . This within one symbol period violates our assumptions and leads
indicates that the correlation between the samples is high when to certain performance degradation. This effect is analyzed in
their distance is close to a multiple of (N/Mi ), and it decays Section V with simulation results.
quickly as the distance to the multiple of (N/Mi ) increases. The denominator of the component timing metric defined
We know the samples whose distance is (d − τ [l2 ] + τ [l1 ]) by (12) can be expanded as
where l1 = l2 are weakly correlated, the summation of their
A −1
l 
N −1
products weighted by zero-mean uncorrelated channel gains
R[n̄i , 0] = |r[n̄i + k]|2 + |r[n̄i + k]|2 = χA + χB .
must be small. Hence, we neglect the second summation in
k=0 k=lA
(41) and replace the first one with (43), (45)

Lh where χA and χB are the summations of the squared mag-
Np nitude of the samples in the former and latter symbols re-
x̂∗ [n̂i + k]x̂[n̂i + k + d] ≈ |Xi [1]|2 ρi (d) φi (d) |h[l1 ]|2 .
N spectively, so they follow a Chi-square distribution with (2lA )
l1 =1
(44) and (2lB ) degrees of freedom scaled by σA 2
/2 and ησA2
/2
Substitute the summation of signal products in (40) with (44), respectively. When lA is sufficiently large, χA can be further
drop the summation of uncorrelated noise products, and use approximated by a Gaussian random variable whose mean and
the definition of σi2 in (4), the approximation to R[n̂i , d] is variance are given by
attained and shown in (6).
A −1
l
 
E {χA } = E |r[n̄i + k]|2 = lA σA
2
, (46)
A PPENDIX B k=0
S TATISTICAL P ROPERTIES OF THE C OMPONENT T IMING
and
M ETRIC AT I NCORRECT T IMING P OSITIONS
  2
We derive the statistical properties of the component timing V ar {χA } = E χ2A − (E {χA })
A −1 l
l A −1
metric in an AWGN channel. Following [16], we approximate  
the received samples at incorrect timing positions as indepen- = E |r[n̄i + k1 ]|2 |r[n̄i + k2 ]|2 − (lA σA
2 2
)
dent complex-valued zero-mean Gaussian random variables. k1 =0 k2 =0
 2
2  2
2
As pointed out in [17], the samples are correlated when there σA σA 2 2
=3 · (2lA ) + 2lA (2lA − 1) − (lA σA )
are unused subcarriers in the OFDM symbols, however, the 2 2
correlation is negligible and the number of unused subcarriers 4
=lA σA . (47)
is small, which is usually the case in practical OFDM systems.
For instance, in the IEEE 802.16 (WiMAX) standard, the This suggests that
√ the mean of χA equals its standard deviation
unused subcarriers (guard bands) occupy about 16% of the multiplied by lA . When lA is large, the deviation of χA
total spectrum. from its mean value is small. Similarly, χB can also be

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
RUAN et al.: TRAINING SYMBOL BASED COARSE TIMING SYNCHRONIZATION IN OFDM SYSTEMS 2567

approximated by its mean when lB is large. It follows that However, lAA , lAB , lBB change with lb differently in various
scenarios. When d ≤ N/2,
2
R[n̄i , 0] ≈ E {χA } + E {χB } = (lA + ηlB ) σA . (48)  
E |R[n̄i , d]|2 |lb = σA 4

If not both lA and lB are large, since (lA + lB = N ), at ⎧

⎪
⎪ (N − d),
most one of them can be small. Without loss of generality, we ⎪
⎪
⎪
⎪ lb ∈ [0, Ng );
assume that lA is small. If η is large, χB dominates the value ⎪
⎪
⎪
⎪ (N − d + N g − lb ) + η(lb − Ng ),
of R[n̄i , 0], then (48) still holds. However, when η is much ⎨
lb ∈ [Ng , Ng + d);
smaller than lA /lB , χA dominates the value of R[n̄i , 0], and · (53)
⎪
⎪ (N − d + Ng − lb ) + ηd + η 2 (lb − d − Ng ),
(48) can be violated. When this happens, we can show that ⎪
⎪
⎪
⎪ lb ∈ [Ng + d, Ng + N − d);
the value of the component timing metric is proportional to ⎪
⎪
√ ⎪
⎪ η(N + N g − l b ) + η 2 (lb − Ng − d),
η, which must be very small and unlikely to cause any false ⎩
lb ∈ [Ng + N − d, Ng + N ).
alarms. This means that (48) is a good enough approximation,
and we can write When d > (N/2),
N 1 |R[n̄i , d]|  
Tc [n̄i , d] ≈ E |R[n̄i , d]|2 |lb = σA4
N − d ρi (d) E {R[n̄i , 0]} ⎧
|R[n̄i , d]| ⎪
⎪ (N − d),
N 1
(49) ⎪
⎪
= 2 . ⎪
⎪ lb ∈ [0, Ng );
N − d ρi (d) (lA + ηlB )σA ⎪
⎪
⎪
⎪ (N − d + Ng − lb ) + η(lb − Ng ),
⎨
|R[n̄i , d]| is the magnitude of the summation of a large number lb ∈ [Ng , Ng + N − d);
· (54)
of zero-mean uncorrelated complex values, so it follows ⎪
⎪ η (N − d),
⎪
⎪
Rayleigh distribution. Tc [n̄i , d] equals |R[n̄i , d]| multiplied by ⎪
⎪ lb ∈ [Ng + N − d, Ng + d);
⎪
⎪
a scaling factor, so approximately follows Rayleigh distribu- ⎪ η(N + Ng − lb ) + η 2 (lb − Ng − d),
⎪
⎩
tion as well. The statistical property of Tc [n̄i , d] is determined lb ∈ [Ng + d, Ng + N ).
by the second moment given by
Substitute the expectation in the integral of (50) with (53)
 
N +Ng −1
   or (54), and after some algebraic manipulation, we obtain
1
E Tc2 [n̄i , d] = E Tc2 [n̄i , d]|lb
N + Ng  2  1 1 N2
 N +Ng
lb =0 E T [n̄ i , d] ≈ · ·
 
c
N + Ng ρ2i (d) (N − d)2
1  
≈ E Tc2 [n̄i , d]|lb dlb (N − d)Ng N −d
N + Ng 0 · − + g(η) (55)
N2 N
1 N2 1
≈ · ·


N + Ng (N − d)2 ρ2i (d) 1 η(η(N −d)+d)
 N +Ng   where g(η)  log (N
Nη
−d)+ηd + log (N −d)+ηd .
E |R[n̄i , d]|2 |lb η−1
· Equation (55) holds for both (d ≤ N/2) and (d > N/2)
2 4 dlb . (50)
0 ((N + (η − 1) max(0, lb − Ng )) σA scenarios. It is easy to compute g(η) = 2(1 − d/N ) as η → 1,
and g(η) = log(N/d) as η → 0 or +∞. Numerical results
For a given lb , we can write show that g(η) is bounded by these limits and only fluctuates
n̄ +N −d−1  within a small range for all d ≥ N/8. Considering the fact
i 
 
E |R[n̄i , d]|2 |lb = E |r∗ [n1 ]r[n1 + d]|2 |lb that η is more likely to fluctuate around 1, it will not introduce
n1 =n̄i much
 2 error to approximate g(η) with g(1) in (55), and
⎧ ⎫ 2
⎨ i  +N −d−1 ⎬ E T [n̄ , d] is approximately equal to (ρ (d)(N − d))−1 .
n̄  c i i
+E r[n1 ]r∗ [n2 ]r∗ [n1 + d]r[n2 + d]|lb
⎩ ⎭
n1 =n̄i n2 =n1
A PPENDIX C
n̄i +N −d−1
    FALSE A LARM P ROBABILITY OF THE COMBINED TIMING
= E |r[n1 ]|2 |lb E |r[n1 + d]|2 |lb (51) METRIC
n1 =n̄i
We index all the component timing metrics from 1 to K.
where the last equality follows the assumption on the samples’
Suppose the k th re-indexed timing metric corresponds to the
independent zero-mean Gaussian distribution. For each indi-
mth component timing metric of training symbol i, we define
vidual sample, lb determines whether it belongs
 to the former 2 2 n
bk  ωi,m σ̄i,m , and b̃n  k=1 bk . We follow the inde-
or latter OFDM symbol, where E |r[n1 ]|2 |lb equals σA 2
and
2 pendent Gaussian distribution assumption about the samples
ησA respectively. Define lAA , lAB , lBB as the numbers of
made in Appendix B, and can show that the component timing
terms in the summation of (51) that are the products of samples
metrics are uncorrelated to each other at incorrect timing
both from the former symbol, one from the former one from
positions. Define
the latter, and both from the latter symbol, respectively. Then,
(51) can be simplified to k 

    4 qk (λ)  P Tc [n̄, di ] > λ , (56)
E |R[n̄i , d]|2 |lb = lAA + η lAB + η 2 lBB σA . (52) i=1

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
2568 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 8, NO. 5, MAY 2009

which can be recursively evaluated by The variance of ZC [n̂1 ] is given by

2
 # $ 2
λ
2 V ar {ZC [n̂1 ]} = λ2 E R̃[n̂i , 0]2 +  K1
f (u1 |b1 )du1 = e−λ /(2b1 ) ,
q1 (λ) = 1 − (57) ρ21 (d1,k )
0 k=1
 +∞  k


K1 
K1 # $
qk+1 (λ) = P Tc [n̄, di ] > (λ − uk+1 ) · |ρ1 (d1,k1 )ρ1 (d1,k2 )|E R̃[n̂1 , d1,k1 ]R̃[n̂1 , d1,k2 ]
0 i=1 k1 =1 k2 =1

· f (uk+1 |bk+1 ) duk+1 2 
K1
 λ − 2λ  K1 |ρ1 (d1,k1 )|
= qk (λ − uk+1 ) f (uk+1 |bk+1 ) duk+1 . (58) k=1ρ21 (d1,k ) k=1
# $
0
· E R̃[n̂1 , d1,k ]R̃[n̂1 , 0] (63)
Assume where we define γi (d) as the sign of ρi (d) and
k−1 −λ2 /(2b̃k )
# $
qk (λ) ≈ Ak λ e (59) R̃[n̂i , 0] = R̃[n̂i , 0] − E R̃[n̂i , 0]| ≈ Θ(n̂i , 0), (64)
# $
where R̃[n̂i , d] = |R̃[n̂i , d]| − E |R̃[n̂i , d]|

√
 12 −k !
k " ≈ γi (d)
(φ∗i (d) Θ(n̂i , d)) . (65)
Ak  ( 2π)k−1 b̃k bi . (60)
For all d1 ≤ d2 , we can show that
i=1
E {Θ∗ (n̂i , d1 )Θ(n̂i , d2 )} = 2σi2 σw
2
φi (d2 − d1 )ρi (d2 − d1 )
When k = 1, the right hand side of (59) reduces to that of · (N − d2 ) (66)
(57), so (59) holds. For k ≥ 1, assume (59) holds for qk . In
the following, we show qk+1 (λ) also satisfies (59) using the E {Θ(n̂i , d1 )Θ(n̂i , d2 )} = 2σi2 σw
2
φi (d2 + d1 )ρi (d2 + d1 )
recursive relationship (58). · max(N − d1 − d2 , 0). (67)
It immediately follows that
qk+1 (λ) # $
 λ E R̃[n̂i , 0]|2 ≈ 2N σi2 σw 2
(68)
2 u 2
/(2b2k+1 ) # $
≈ Ak (λ − u)k−1 e−(λ−u) /(2b̃k ) e−u du
0 bk+1 E R̃[n̂i , d]R̃[n̂i , 0] ≈ (N − d) |ρi (d)| σi2

2  
· N (σi2 + σw 2 2
bk+1
u− λ
) + 2σw (69)
# $
b̃k+1
 λ
− bk+1 b̃k

=
Ak −λ2 /(2b̃k+1 )
e u(λ − u)k−1 e
2
b̃k+1
du E R̃[n̂i , d1 ]R̃[n̂i , d2 ] ≈ γi (d1 )γi (d2 ) ((N − d2 )
bk+1 0

2 · ρi (d2 − d1 ) + ρi (d2 + d1 )
bk+1
u−
b̃k+1
λ
· max(N − d2 − d1 , 0)) . (70)
 k−1  +∞
− bk+1 b̃k
Ak −λ2 /(2b̃k+1 ) bk 2
Thus, V ar {ZC [n̂1 ]} can be obtained by substituting (68), (69)
≈ e k
λ e b̃k+1
du
b̃k+1 b̃k+1 −∞ and (70) into (63), and the result is shown in (38). Let all
2
=Ak+1 λk e−λ /(2b̃k+1 )
(61) ρi (d) = 1 in (38), we can simplify the expression for the
variance of ZA [n̂1 ] to that in (26).
where the first approximation follows the assumption about
A PPENDIX E
qk (λ), the second approximation follows the fact that the
T HE M EAN AND VARIANCE OF ZB [n̂1 ]
exponential term approximates 0 very quickly out of the
b
small region around u = b̃k+1 λ. Therefore, (59) holds for The mean of ZB [n̂1 ] can be directly evaluated by
k+1
all k. Combining it with (56), (17) is obtained. The derivation E {ZB [n̂1 ]} = N 2 σ12 (σ12 + σw
2
) − λN 2 (σ12 + σw
2 2
) . (71)
indicates that the approximation is accurate for large λ values.
The second moment of ZB [n̂1 ] can be expanded to
     
E ZB [n̂1 ]2 = λ2 E R[n̂1 , 0]2 E R[n̂2 , 0]2
# $
A PPENDIX D + E (|RB,1 [n̂1 ]|R[n̂2 , 0] + |RB,2 [n̂2 ]|R[n̂1 , 0])
2
T HE M EAN AND VARIANCE OF ZA [n̂1 ] AND ZC [n̂1 ]  
− λE {|RB,1 [n̂1 ]|R[n̂1 , 0]} E |RB,2 [n̂2 ]|2
 
Because Case A can be seen as a special case of Case C, the − λE {|RB,2 [n̂2 ]|R[n̂2 , 0]} E |RB,1 [n̂1 ]|2 . (72)
mean and variance of ZA [n̂1 ] can be obtained by simplifying According to the definition of RB,i [n̂i ], we can compute
those of ZC [n̂1 ] with an extra condition (ρ1,k = 1). The mean
of ZC [n̂1 ] can be directly computed from (36) as E {RB,i [n̂i ]R[n̂i , 0]}
  Mi − 1
=N σi2 N (σi2 + σw 2 2
) + 2σw . (73)
E {ZC [n̂1 ]} = N σ12 − λN (σ12 + 2
σw ). (62) M1 + M2 − 2

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.
RUAN et al.: TRAINING SYMBOL BASED COARSE TIMING SYNCHRONIZATION IN OFDM SYSTEMS 2569

Also, [12] T. Bhatt, V. Sundaramurthy, J. Zhang, and D. McCain, “Initial syn-

  chronization for 802.16e downlink," in Proc. Fortieth Asilomar Conf.
E |RB,i [n̂i ]|2 Signals, Systems Computers (ACSSC), Oct. 2006.
 2 M i −1 M
i −1
[13] IEEE LAN/MAN Standards Committee, “Channel models for fixed
2 wireless applications," Document IEEE 802.16.3c-01/29r4, 2003.
= E {|R[n̂i , di,k1 ]R[n̂i , di,k2 ]|}
M1 + M2 − 2 [14] ITU-R M.1225, “Guidelines for evaluation of radio transmission tech-
k1 =1 k2 =1
 2 nologies for IMT-2000," Recommendation ITU-R M 1225, 1997.
Mi − 1 [15] W. C. Jakes Jr. (ed.), Microwave Mobile Communications. J. Wiley &
=N σi2 (N σi2 + 2σw 2
) . (74) Sons, 1974.
M1 + M2 − 2 [16] Jan-Jaap van de Beek, M. Sandell, and P. O. Börjesson, “ML estimation
Thus, ignoring the higher order noise terms, we can approxi- of time and frequency offset in OFDM systems," IEEE Trans. Signal
Processing, vol. 45, pp. 1800-1805, July 1997.
mate the products of expectations as [17] S. H. Müller-Weinfurtner, “On the optimality of metrics for coarse
# $
2 frame synchronization in OFDM: a comparison," in Proc. Ninth IEEE
E (RB,1 [n̂1 ]R[n̂2 , 0] + RB,2 [n̂2 ]R[n̂1 , 0]) International Symposium on Personal, Indoor and Mobile Radio Com-
munications, vol. 2, Sept. 1998, pp. 533-537.
≈N 4 σ16 (σ12 + 2σw
2
) + 4N 3 σ16 σw
2
(75)
 2
  2

E R[n̂1 , 0] E R[n̂2 , 0] Ming (Matt) Ruan received the B.Eng and M.Eng
degrees from Shanghai Jiaotong University, Shang-
≈N 4 σ16 (σ12 + 2
4σw ) + 4N 3 σ16 σw
2
(76) hai, China, in 1998 and 2001, respectively. Then he
 2  joined Lucent Bell Labs China as a member of tech-
E {RB,1 [n̂1 ]R[n̂1 , 0]} E R [n̂2 , 0] nical staff. In 2005 he immigrated to Australia and
  M1 − 1 became a PhD candidate in the Australian National
≈ N 4 σ16 (σ12 + 3σw2
) + 4N 3 σ16 σw
2
(77) University with the support of an Australian Post-
M1 + M2 − 2
 2  graduate Award (APA) and a supplemental schol-
E {RB,2 [n̂2 ]R[n̂2 , 0]} E R [n̂1 , 0] arship from the National ICT Australia (NICTA).
  M2 − 1 Since 2008 he has been working at NICTA as a
≈ N 4 σ16 (σ12 + 3σw2
) + 4N 3 σ16 σw
2
. (78) research engineer. His current research interests in-
M1 + M2 − 2 clude synchronization technologies, statistical signal processing, and iterative
Combine them with (72), then minus the square of the mean receiver design.
given by (71), (32) is attained. Mark C. Reed has worked in industry and research
positions for the last 18 years with positions in the
R EFERENCES U.S.A., Switzerland, and Australia. He pioneered
[1] T. M. Schmidl and D. C. Cox, “Robust frequency and timing synchro- the area of iterative receiver design as part of his
nization for OFDM," IEEE Trans. Commun., vol. 45, pp. 1613-1621, doctoral studies and was part of a team that de-
Dec. 1997. signed and developed a world first Satellite-UMTS
[2] A. J. Coulson, “Maximum likelihood synchronization for OFDM using Modem for the European Space Agency. He also
a pilot symbol: algorithm," IEEE J. Select. Areas Commun., vol. 19, pp. completed further work on 3G Basestation design
2486-2494, Dec. 2001. as technical lead in the highly successful European
[3] ——, “Maximum likelihood synchronization for OFDM using a pilot Commission project, ASILUM, which investigated
symbol: analysis," IEEE J. Select. Areas Commun., vol. 19, pp. 2495- and validated advanced signal processing schemes
2503, Dec. 2001. for link improvement in UMTS. Since April 2003 Dr Reed is employed
[4] H. Minn, V. K. Bhargava, and K. B. Letaief, “A robust timing and as a Principal Researcher at National ICT Australia (NICTA), and is an
frequency synchronization for OFDM systems," IEEE Trans. Wireless Adjunct Associate Professor at the Australian National University, Canberra,
Commun., vol. 2, July 2003. Australia, where he is involved in research, education, and commercialization,
[5] K. Shi and E. Serpedin, “Coarse frame and carrier synchronization of within the wireless signal processing program. He has over 45 international
OFDM systems: a new metric and comparison," IEEE Trans. Wireless journal and conference papers, was previously an Associate Editor for the
Commun., vol. 3, July 2004. IEEE T RANSACTIONS ON V EHICULAR T ECHNOLOGY, and has been listed
[6] Y. Li, J. L. J. Cimini, and N. R. Sollenberger, “Robust channel estimation as inventor on eight patent applications.
for OFDM systems with rapid dispersive fading channels," IEEE Trans.
Commun., vol. 46, pp. 902-915, July 1998. Zhenning Shi received the BS degree in Electronics
[7] E. G. Larsson, G. Liu, J. Li, and G. B. Giannakis, “Joint symbol timing Engineering from Tsinghua University, China, in
and channel estimation for OFDM-based WLANs," IEEE Commun. 1998, and the PhD degree in Electrical and Com-
Lett., vol. 5, pp. 325-327, Aug. 2001. puter Engineering from the University of Utah, US,
[8] R. A. Pacheco, O. Üreten, D. Hatzinakos, and N. Serinken, “Bayesian in 2003. From 2002 to 2004, he was a post-doctoral
frame synchronization using periodic preamble for OFDM-based fellow at the Electrical and Computer Engineering
WLANs," IEEE Signal Processing Lett., vol. 12, pp. 524-527, July 2005. Department, the University of Utah. In 2004, he
[9] H. Minn, V. K. Bhargava, and K. B. Letaief, “A combined timing and joined NICTA and has been with the Canberra
frequency synchronization and channel estimation for OFDM," IEEE Research Lab since then. He is also an adjunct
Trans. Commun., vol. 54, pp. 416-422, Mar. 2006. research fellow at Australian National University
[10] IEEE, Par 11: Wireless LAN medium access control (MAC) and (ANU), where he is involved in the joint education
physical layer (PHY) specifications: high-speed physical layer in the and research program in wireless communications. Dr. Zhenning Shi’s primary
5 GHz band. IEEE-SA Standards Board, 2003. research interests include multiuser detection, iterative receivers, channel
[11] ——, IEEE Standard for local and metropolitan area networks, estimation, equalization and synchronization in CDMA, OFDMA and LTE.
part 16: air interface for fixed mobile broadband wireless He has published over 50 technical papers and filed four patent applications
access systems, amendment 2: physical and medium access in these areas.
control layers for combined fixed and mobile operation
in licensed bands and corrigendum 1. [Online] Available:
http://ieeexplore.ieee.org/servlet/opac?punumber=10676, 2006.

Authorized licensed use limited to: VELLORE INSTITUTE OF TECHNOLOGY. Downloaded on July 28, 2009 at 06:34 from IEEE Xplore. Restrictions apply.